Skip to content
Beyond Prompt AI Studio
Comparing AI models properly

Why good test scores can deceive: contamination and the training-data curse

A high benchmark score looks impressive – as long as you don't know whether the model already saw the test questions during training. Contamination is exactly that problem, and it affects practically every widely used benchmark to some degree.

Assessment as of: July 2026 – research moves fast, this assessment may change.

Four examples – worth remembering

Try it yourself: match the contamination fact to its number

Fact

Documented finding

How test data ends up in training

Language models are trained on vast amounts of internet text. Publicly available benchmarks can be freely downloaded – and, often unintentionally, end up in exactly that training data. A model that has already seen the test questions along with their correct answers can recall them rather than actually solve them.

How big the problem actually is

A study by Johns Hopkins University found signs of contamination in 29.1% of MMLU test items. Other audits found leakage rates ranging from 1% to 45% depending on the benchmark. Some studies of multilingual benchmarks measured even higher contamination rates – a sign that the problem tends to grow over time rather than shrink, since newly released benchmarks spread quickly through forums, copies, and derivative datasets.

A concrete before/after example

When contaminated test items are swapped for an unpublished, "clean" variant of equal difficulty, results often drop noticeably: in one such study, a Mistral model's score fell by up to 13 percentage points on a decontaminated GSM8K variant compared to the original result.

Why rewording doesn't reliably evade detection

Contamination isn't limited to verbatim copying. A paraphrased or translated version of a test question can also evade standard detection methods, even though it's substantively the same task as the original – and still produce an artificially inflated score.

Why this matters to you as a decision-maker

A single high benchmark score alone doesn't tell you whether a model truly has the underlying capability or already knows the test questions. Results on privately held, current, or purpose-built test questions (more on this in this course's final module) are far more meaningful than a score on a benchmark that's been publicly available for years.

Key takeaways

  • Contamination happens when publicly accessible test data unintentionally ends up in a model's training data.
  • One study found contamination signs in 29.1% of MMLU test items; other audits measured 1–45% leakage depending on the benchmark.
  • A Mistral model lost up to 13 percentage points on a decontaminated GSM8K variant – direct evidence of contamination's effect on a measured score.
  • Contamination isn't limited to verbatim copying – paraphrased or translated test questions can still evade detection methods.
  • Results on private, current test questions are more meaningful than a score on a benchmark that's been publicly available for years.

Quick check: did it land?

1 / 3

According to this module, how does contamination arise in the vast majority of cases?

Curious whether a benchmark score really shows what it claims?